PASD: Partner-Aware Skill Discovery for Human-AI Collaboration
A new framework called Partner-Aware Skill Discovery (PASD) has been introduced by researchers, utilizing deep hierarchical reinforcement learning to enhance collaboration between humans and AI by considering partner behavior during skill acquisition. Traditional DHRL approaches prioritize agent-focused rewards, which can lead to shortcut learning, where skills rely on misleading information instead of adjusting to the changing behaviors of partners. PASD incorporates a contrastive intrinsic reward that identifies patterns from interactions with partners, ensuring that skill representations are aligned among similar partners while remaining distinct across various strategies. This method reduces shortcut learning and fosters effective adaptation to new partners. The findings are available on arXiv with the identifier 2605.24352.
Key facts
- PASD is a DHRL framework for human-AI collaboration.
- It learns skills conditioned on partner behavior.
- Contrastive intrinsic reward captures partner interaction patterns.
- Aligns skill representations across similar partners.
- Maintains discriminability across diverse partner strategies.
- Mitigates shortcut learning in multi-agent collaboration.
- Published on arXiv:2605.24352.
- Addresses adaptation to novel partners with dynamic behaviors.
Entities
Institutions
- arXiv